IP Hot Topic: Incident to the Very Nature of Machine Learning – Federal Circuit Says Training Data Alone Won’t Rescue AI Patent Claims

On July 7, 2026, the Federal Circuit affirmed summary judgment that claims of two machine learning patents covering orthodontic image analysis are invalid under 35 U.S.C. § 101. Dental Monitoring SAS v. Align Technology, Inc., No. 2024-2270 (Fed. Cir. July 7, 2026) (nonprecedential). The decision is an early application and extension of Recentive Analytics, Inc. v. Fox Corp., 134 F.4th 1205 (Fed. Cir. 2025) beyond data analytics into image analysis and neural-network-based assessments. The message for AI patent owners and defendants alike: claims that merely apply conventional machine learning techniques to a new application area, without claiming a technological improvement in the machine-learning implementation itself, face heightened eligibility risk.

Background

Dental Monitoring SAS owns U.S. Patent Nos. 11,049,248 and 10,755,409, both directed to dental arch image analysis. The ’248 patent is “directed to a method for assessing the shape of an orthodontic aligner using a ‘deep learning device,’” while the ’409 patent “is directed to a method for acquiring an image of a dental arch and analyzing it using a ‘deep learning device.’” Both methods automate work traditionally performed by dental practitioners. Aligner therapy requires periodic assessment of patient progress to evaluate whether to advance to the next aligner. The two patents propose replacing those in-person assessments with patient-captured photos analyzed by software.

A “deep learning device,” as the patents describe it, is a “a machine learning device that, through training, can analyze images and recognize patterns within the images.” Slip Op. at 2. Critically, the specifications said the device could be chosen from preset lists of well-known image classification and object detection networks, including ones made by Google and Microsoft.

Dental Monitoring sued Align Technology in the Northern District of California, accusing Align’s Invisalign Virtual Care AI platform of infringement. Judge Alsup ordered a “patent showdown”: each side selected one claim, took targeted discovery, and filed cross-motions for summary judgment, with the ruling to govern three related claims by stipulation. The district court held the claims ineligible, and Dental Monitoring appealed.

The Federal Circuit’s Analysis

The Recentive baseline. In Recentive, the Federal Circuit held that claims doing no more than applying established machine learning methods to a new data environment are not patent eligible; iterative training and dynamic updating are “incident to the very nature of machine learning” and supply no technological improvement. The court there, like here, applied the Alice two-step framework to assess patent eligibility for an AI-based invention.

Alice step one. Writing for the panel, Judge Lourie held both claims directed to abstract ideas: claim 14 of the ’248 patent to “collecting and analyzing image information using a deep learning device,” and claim 12 of the ’409 patent to acquiring an image, analyzing it via a “deep learning device,” comparing the result to a setpoint, and transmitting the result. Both fall within the “familiar class” of ineligible claims recognized in Electric Power Group, LLC v. Alstom S.A., 830 F.3d 1350 (Fed. Cir. 2016): collecting information, analyzing it, and displaying the results.

The “deep learning device” changed nothing. Citing Recentive, the panel held that the claims simply applied generic machine learning to a new environment: “dental arch image analysis.” The requirement that the device be trained on more than 1,000 labeled dental arch images fared no better; the court treated training a model on a specific subset of domain-relevant data as a conventional aspect of machine learning rather than a technological improvement. The decision does not suggest that training-data limitations lack patentable significance in all contexts. Rather, the court viewed the particular training limitations here as conventional inputs to a generic machine-learning framework and therefore insufficient to supply eligibility absent a claimed technological improvement. And Dental Monitoring’s argument that the method assesses aligner fit more precisely than a human failed on the claim language: nothing in claim 14 requires assessment beyond what orthodontists already do. The panel emphasized that greater speed and efficiency alone cannot confer eligibility.

Alice step two. The panel found no inventive concept. The patents themselves conceded that the deep learning device was conventional by listing off-the-shelf neural networks. That the combination as a whole may have been unconventional missed the mark because “the relevant inquiry is not whether the claimed invention as a whole is unconventional or non-routine.” The specification’s admissions also defeated the fallback argument that factual disputes about conventionality precluded summary judgment.

The summary judgment posture matters. Recentive was decided on the pleadings. This case reached the Federal Circuit on summary judgment after targeted discovery, with an identical outcome. Patentees often invoke Berkheimer to argue that conventionality is a fact question for trial. Here the intrinsic record answered the question; discovery gave the patentee no refuge.

What the court did not decide. The panel did not say that machine learning patents are categorically ineligible or that computer-vision inventions are categorically abstract. Nor did the panel find that trained neural networks are per se conventional. Rather, the court focused on generic neural networks listed in the specification, conventional training, and a lack of claimed technological improvement.

McRO remains an important counterpoint. The decision does not displace cases such as McRO, Inc. v. Bandai Namco Games America Inc., 837 F.3d 1299 (Fed. Cir. 2016). There, the Federal Circuit upheld claims automating 3-D lip-synchronization animation because they recited rules with specific, claimed characteristics: morph weights defined as a function of phoneme sub-sequences and timing. Those rules performed a process different from the subjective, judgment-driven method animators had used, and their claimed structure avoided preempting other rules-based approaches. Recentive distinguished McRO on this ground, and Dental Monitoring’s claims fall on the wrong side of the line: the deep learning device merely performed the same assessment orthodontists already perform, using generic machine learning techniques and without claiming the purported technological improvement itself. Automation claims are most likely to survive § 101 when they recite a specific technical mechanism that changes how the computer performs the task, rather than merely automating a task previously performed by humans. These did not.

Five Takeaways for In-House Counsel

1. Recentive applies more broadly. A different panel applied Recentive without hesitation to computer vision in a clinical setting. The “new data environment” rule is not limited to the scheduling or analytics of Recentive.
2. Domain-specific training data alone is unlikely to save a claim. Requiring training on 1,000-plus labeled, field-specific images was still treated as “incident to the very nature of machine learning.” If your portfolio relies on training-regimen limitations for eligibility, reassess it.
3. Beware of comparisons to commercially available neural networks. If the invention resides in the model, do not characterize the model as interchangeable with commercially available neural networks absent a clear explanation of what is technologically different.
4. McRO remains the playbook. Claims are most likely to survive § 101 when they recite the specific technical means that produces the asserted improvement. McRO survived because the claims required particular rule structures that changed how the computer performed animation relative to the process human animators used. Dental Monitoring failed because the claims relied on generic machine-learning tools and did not claim the alleged technological advance itself.
5. Section 101 works at any stage. Accused infringers now have affirmed wins on the pleadings (Recentive) and at summary judgment (this case). Consider streamlined procedures like Judge Alsup’s “showdown” to force early eligibility rulings.

One caution: the decision is nonprecedential and binds no future panel. Nevertheless, its unqualified application of Recentive suggests that Recentive’s reasoning may continue to influence the Federal Circuit’s treatment of AI-related eligibility challenges.